metadata/STATISTICS.md

LogicBench-1K — Dataset Statistics (camera-ready)

Version: 1.0.1
Generated: 2026-06-08T19:05:59Z
Samples: 1,000 (700/100/200 train/val/test)

Sample diversity

Metric	Value
Unique gate mixtures	343
Mixture coverage (unique / samples)	34.3%
Gates per diagram (mean ± std)	6.174 ± 3.699
Gates per diagram (median [Q1, Q3])	5.0 [4.0, 7.0]
Gates per diagram (range)	1–31
Diagrams with no edges	19

Node-count histogram

Bin (gates)	Count
1-5	576
6-10	343
11-15	49
16-20	18
21-31	14

Topology distribution

Gate-type frequencies (global)

Gate	Count	Share
AND	1,492	24.2%
NAND	915	14.8%
NOR	597	9.7%
NOT	1,389	22.5%
OR	894	14.5%
XNOR	258	4.2%
XOR	629	10.2%

Edge-count summary

Statistic	Value
Total directed edges	6,423
Edges per diagram (mean ± std)	6.423 ± 5.636
Edges per diagram (median)	5.0
Edges per diagram (range)	0–57
Mean directed density	0.2355

Edge-count histogram

Bin (edges)	Count
0	19
1-5	544
6-10	298
11-20	114
21-57	25

Split-stratified summary

Split	n	Unique mixtures	Gates (mean)	Edges (mean)	Density
Train	700	261	5.91	6.08	0.2399
Val	100	73	6.77	7.04	0.2148
Test	200	128	6.8	7.315	0.2305

Paper-ready summary

The corpus comprises 1,000 annotated diagrams with 6,174 gates and 6,423 directed edges. Sample diversity spans 343 unique gate mixtures (34.3% coverage). Diagrams contain 1–31 gates (mean 6.174, median 5.0) and 0–57 edges (mean 6.423, median 5.0). Gate-type and topology statistics are stratified by the official 700/100/200 train/validation/test split in metadata/statistics.json.

---

Auto-generated by scripts/generate_statistics.py. Machine-readable source: statistics.json.